Route Simple Tasks to Qwen / DeepSeek to Reduce Costs

Choose Qwen or DeepSeek in your request or client for summarization, classification, rewriting, and other simple tasks.

The value you directly control is model. For simple tasks, set model to a Qwen or DeepSeek model. For complex tasks, use OpenAI, Claude, or another stronger model.

Step 1: Confirm an available model name

Open Model Plaza, find a Qwen or DeepSeek model available to your account, and copy the full model name. You can also check the model list API:

curl https://api.routescope.ai/v1/models \
  -H "Authorization: Bearer sk-your-token"

Replace your-qwen-model, your-deepseek-model, and your-strong-model in the examples below with real model names available to your account.

Step 2: Send simple tasks to Qwen / DeepSeek

For summarization, classification, format conversion, and short rewrites, change the request model to Qwen or DeepSeek.

curl https://api.routescope.ai/v1/chat/completions \
  -H "Authorization: Bearer sk-your-token" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "your-qwen-model",
    "messages": [
      {
        "role": "user",
        "content": "Summarize the following text into three bullet points: ..."
      }
    ],
    "max_tokens": 400
  }'

If you use Codex, VS Code / Cline, OpenCode, or another OpenAI-compatible client, fill the Qwen or DeepSeek model name in the client model setting. The Base URL stays:

https://api.routescope.ai/v1

Step 3: Pick a model by task type in your code

If you can edit your application code, choose the model based on task type.

const SIMPLE_TASK_MODELS: Record<string, string> = {
  summarize: "your-qwen-model",
  classify: "your-qwen-model",
  rewrite: "your-deepseek-model",
  translate: "your-qwen-model",
};

function pickModel(taskType: string) {
  return SIMPLE_TASK_MODELS[taskType] ?? "your-strong-model";
}

const completion = await client.chat.completions.create({
  model: pickModel(taskType),
  messages,
  max_tokens: taskType === "classify" ? 200 : 1000,
});

Recommended choices

Task	Recommended model direction	Configuration tip
Classification, tags, intent detection	Qwen / DeepSeek	Use a lower-cost model and constrain the output format.
Summarization, rewriting, translation	Qwen / DeepSeek	Control `max_tokens` and test quality first.
Batch data cleaning	Qwen / DeepSeek	Process in batches and watch average usage in operation records.
Complex code generation	OpenAI / Claude / stronger reasoning models	Do not optimize only for cost; check quality and failure rate.
Critical production responses	Strong model, plus backup model if needed	Add fallback logic in your own code.

Cost-saving tips

For classification and extraction, ask for only the required format, such as JSON with no explanation.
Set a smaller max_tokens for simple tasks.
Test 20 to 50 real samples before scaling a batch job.
Do not guess model names. Copy them from Model Plaza or /v1/models.

Verify the savings

After a test request, open operation records and check:

Item	What to check
Requested model	It is the Qwen or DeepSeek model you selected.
Status	The request succeeded, or the error is clear.
Tokens	Input and output tokens match expectations.
Cost	It is lower than the same task on a stronger model.